A. Junghanns and J. Schaeffer. Domain-dependent single-agent search enhancements. In Proc of IJCAI-99, pages 570--575, Stockholm, Sweden, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....h(n) might visit every little town in the peninsula before trying to go around. KBFS will also explore nodes outside that subtree and therefore may reach the goal faster. A special case of a dead end is a deadlock, which is a dead end that cannot be recovered from. For example, the game of Sokoban[10] contains numerous deadlocks. Consider the graph of Figure 3 where the value written next to a node represents its cost value. The two subtrees rooted at B and C are structurally symmetric, but node B is the root of a dead end subtree. KBFS(1) i.e. BFS, might expand the nodes in the following ....

A. Junghanns and J. Schaeffer. Domain-dependent single-agent search enhancements. In Proc of IJCAI-99, pages 570--575, Stockholm, Sweden, 1999.

.... Men s Morris [43] Connect 4 [103] Qubic [104] and Go Moku [104] This chapter has not addressed one player games (or puzzles) Single agent search has been successfully used to optimally solve the 15 puzzle [14] and Rubik s Cube [105] and progress is being made on solving Sokoban problems [106]. Recently, major advances have occurred in building programs that can solve crossword puzzles [107] The last few years have seen research on team games become popular. The annual RoboCup competition encourages hardware builders and software designers to test their skills on the soccer eld ....

A. Junghanns and J. Schaeer. Domain-dependent single-agent search enhancements. In International Joint Conference on Articial Intelligence, pages 570-575, 1999.

....impossible with current technology to search deep enough with standard full width search to get a high performance program. Examples are games with a large average number of legal moves like Go and shogi [14] and single agent search problems with extremely long solution sequences such as sokoban [10]. To make a high performance program in these domains, some method for plausible move generation is needed [6, 24, 12, 26, 11] Especially in Go, most of the available time per move is spent on generating promising looking moves, leaving little time for search [6] Therefore, despite failing to ....

A. Junghanns and J. Schaeer. Domain-Dependent Single-Agent Search Enhancements. In Proceedings of the Sixteenth International Joint Conference on Articial Intelligence (IJCAI-99), pages 570-575, 1999.

....the heuristic function is given by the user, but we expect to be able to exploit the richer language for extracting better heuristics. Except for the modeling point, the rest of the issues are common to the problems encountered in the application of domain specific heuristic search methods [JS99]. Indeed, the only thing that distinguishes planning as heuristic search from classical heuristic search is the use of a general declarative language for encoding problems and getting the heuristic information. The biggest challenge is to make the declarative approach competitive with specialized ....

A. Junghanns and J. Schaeffer. Domain-dependent single-agent search enhancements. In Proc. IJCAI-99. Morgan Kaufmann, 1999.

....problems and domains, and introduce a family of planners that are competitive with some of the best current planners. Planners based on the ideas of heuristic search are related to specialized solvers such as those developed for domains like the 24 puzzle [KT96] Rubik s cube [Kor98] and Sokoban [JS99] but differ from them mainly in the use of a general language for stating problems and a general mechanism for extracting heuristics. Heuristic search planners, as all planners, are general problem solvers in which the same code must be able to process problems from different domains [NS63] This ....

....language for stating problems and a general mechanism for extracting heuristics. Heuristic search planners, as all planners, are general problem solvers in which the same code must be able to process problems from different domains [NS63] This generality comes normally at a price: as noted in [JS99], the performance of the best current planners is still well behind the performance of specialized solvers. Closing this gap, however, is the main challenge in planning research where the ultimate goal is to have systems that combine flexibility and efficiency: flexibility for modeling a wide ....

[Article contains additional citation context not shown here]

A. Junghanns and J. Schaeffer. Domain-dependent single-agent search enhancements. In Proceedings IJCAI-99. Morgan Kaufmann, 1999.

....usable. Adding game speci c knowledge for local move pruning and thereby using a locally informed global search can greatly enhance performance and reduce the size of search trees by many orders of magnitude. Similar experiences have been reported for the single agent search problem of Sokoban [7]. These results point the way towards developing a common framework that combines the power of local search methods with the generality of global minimax search. Such future work includes: Use a combination of local and global search in the case when there are some dependencies between local ....

A. Junghanns and J. Schaeer. Domain-dependent single-agent search enhancements. In IJCAI-99, pages 570-575, 1999.

....However, there are games in which it is impossible with current technology to search deep enough with standard full width search to get a high performance program. Examples are games with a large average number of legal moves like Go and shogi [13] and single agent search problems such as sokoban [9]. Here plausible move generation can be a good alternative to full width search. Plausible move generation is also interesting for cognitive science. Despite the success of the full width search approach, there has been debate about the level of arti cial intelligence used in these programs (a ....

A. Junghanns and J. Schaeer. Domain-Dependent Single-Agent Search Enhancements. In Proceedings of the Sixteenth International Joint Conference on Articial Intelligence (IJCAI-99), pages 570-575, 1999.

....heuristics. These considerations are outside the scope of this paper. Typically the choice of basic algorithm (single two agent) is easily made based on the problem de nition. For most applications, the majority of the design e ort involves judiciously ne tuning the set of algorithm enhancements [11, 12]. The applicability of search algorithm enhancements is determined by the ve categories of properties of the state space. Figure 1 summarizes the interaction between the state space properties (x axis) and step 2 of the algorithm design process the enhancements (the y axis) A sampling of ....

A. Junghanns and J. Schaeer. Domain-dependent single-agent search enhancements. In IJCAI-99, pages 570-575, 1999.

....that make minimax search dicult even for small size problems. On the other hand, adding game speci c knowledge can greatly enhance performance and reduce the size of search trees by many orders of magnitude. Similar experiences have been reported for the single agent search problem of Sokoban [3]. While global minimax search cannot compete with the local search method used in Decomposition Search, adding game speci c knowledge for move pruning and using locally informed global search greatly improves the eciency of minimax searches. This result points the way towards developing a common ....

A. Junghanns and J. Schaeer. Domaindependent single-agent search enhancements. In IJCAI-99, pages 570-575, 1999.

.... 4] Previously, we reported on our attempts to solve Sokoban problems using the standard single agent search techniques available in the literature [10] When these proved inadequate, solving only 10 problems of a 90 problem test suite, new algorithms had to be developed to improve search eciency [8, 9, 11, 12]. This allowed 47 problems to be solved optimally or near optimally. Additional e orts have since increased this number to 57. The results reported here document the large gains achieved by adding application dependent knowledge to our program, Rolling Stone. Many of the search enhancements added ....

....nodes dominate the cost of the search for some problems. Some additional heuristics for deciding when to execute pattern searches could result in further improvements in the overall search eciency. There are numerous parameters in the search, each of which can be tuned for maximal performance [7, 11]. Pattern searches have also been applied to sliding tile puzzles [7] The pro14 gram dynamically learns penalty patterns, such as linear con icts [6] The cost of the pattern searches is small compared to the large reductions in the IDA search tree. Deadlock tables (or pattern databases) are ....

A. Junghanns and J. Schaeer. Domain-dependent single-agent search enhancements. In IJCAI, pages 570-575, 1999.

....without changing the general topology of the problem would change the geographic distance, but not the influence. The following is a list of properties we would like the influence measure to reflect: 3 Subsequent to this work, further refinements have pushed the number of problems solved to 52[8]. Fig. 2. The Number of Alternatives Changes the Influence Fig. 3. The Location of the Goals Matters Alternatives: The more alternatives that exist on a path between two squares, the less the squares influence each other. That is, squares in the middle of a room where stones can go in all 4 ....

A. Junghanns and J. Schaeffer. Domain-dependent single-agent search enhancements. In IJCAI, 1999. To appear.

....on an earlier version, Relevance Cuts: Localizing the Search [JS98a] which was presented in 1998 at The First International Conference on Computers and Games , Tsukuba, Japan. Chapter 8 Single Agent Search Enhancements is based on the paper DomainDependent Single Agent Search Enhancements [JS99a] which was presented at IJCAI 99, Stockholm, Sweden. Chapter 2 Single Agent Search 2.1 Purpose of Search Real world problems can often be abstracted into models where a state of the world is described mathematically. State transition rules describe the conditions for the transitions between ....

A. Junghanns and J. Schaeer. Domain-dependent single-agent search enhancements. In IJCAI, pages 570-575, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC